Density plots

These density plots show the distribution of tempo in Beats Per Minute (BPM) for the movie genres. On average, Feelgood/Comedy movies have the highest BPM. The average BPM for Horror movies is the lowest (89 BPM). The distribution for Romance/Drama and Action is about the same. It is also clear that overall, songs in minor key do have a slightly lower BPM than major-key songs, especially in Action movies.

The expectation was that the Action movie has the highest overall tempo from all genres. An explanation for this result is that the average Action movie has both very uptempo tracks, but also very quiet en slow tracks. Slow ‘Horror-like’ tracks are used to build up tension. That is why the distribution for major Action tracks is quite wide. The tempo distribution of minor songs in Action movies is also a surprising result. Apparently slower songs used in the Action movie are mostly written in a minor key. As you may remember, songs written in a minor key are perceived as more depressing. So slow tracks are being used to convey a ‘depressing’ or ‘negative’ feeling and more uptempo tracks are being used to convey more positive, maybe ‘hero-like’ feelings.

Self-similarity matrices

In a self-similarity matrix each element of the feature sequence is being compared with all other elements. Path-like structures represent exact repetitions. There is one main diagonal visible, this is because both axis represent the exact same song. Block-like structures represent homogeneous regions. This is where music features stay somewhat constant over the duration of an entire musical part.

The left visualizations represent self-similarity matrices for ‘chroma’. It demonstrates at which points in the track the same pitches occur. The right visualizations represent the same song, but with ‘timbre’, also referred to as ‘tone color’. Later on there will be more information about this musical feature.

‘Feelgood’

is a typical Feelgood/Comedy track (what else!) from ‘They Came Together’. The SSM of chroma shows a block-like pattern. At t = 25 some percussion instruments and a piano come to the fore. In the SSM of timbre the last section is very distinguishable. It is clear that a different instrumentation is used. The lead singer stops and a electric guitar starts playing.

‘Another Dance’

is a track from the Romantic/Drama ‘Pride and Prejudice’. This is a good example of exact repetitions. Only diagonal lines are visible, but no big block-like structures. Lines that are diagonal even to the main diagonal are exact repetitions. It is very clear when you listen to this track.

Judge yourself!

Chroma- and keygrams

Both left grams represent chromagrams. These sum up all pitch coefficients that belong to the same chroma, so this gram cyclic in nature. The graphics to the right represent chordograms with all chords used in the track.

Cepstrograms: Timbre components

Timbre, also known as “tone color”, is the perceived sound quality of a musical note or sound. It distinguishes various types of musical instruments. There are twelve timbre coefficients in total. The values are high level abstractions of the spectral surface ordered by degree of importance.

• c01: The ‘average loudness’.
• c02‘brightness’. Increased levels of mid and high frequency content are referred to as ‘brighter’.
• c03: Closely related to the ‘flatness’ of a sound.
• c04: Sounds with a ‘stronger attack’.

Doll box is different than the majority of the horror film music. It has a extremely high valence (0.854) for Horror film music. This track is being compared with Curse Your name, which is a very typical horror track with an extremely low valence (0.021). The energy values are more or less the same (0.0625 and 0.0725). Doll box is written in a major key, with very high pitched, distinctive tones.

In Doll Box it is very clear which timbre components are being used in this track; c02, c03 and c04. The sound of this track represents the typical sound of a doll music box. Because these three components are very constant throughout the whole track, it is hard to distinguish the different sound characteristics. Two clear sections in the cepstrogram are the very contrasting parts at t = 1 and at t = 32. These parts are riffs on a copper xylophone.

As you can see, the timbral components of Curse Your Name are much more spread out in the cepstrogram. This is very typical in horror music, because horror music tends to have a lot of different musical characteristics/instruments played at the same time, that seeks to give the audience an uncomfortable and unsettling feeling. The first bright yellow part of co2 is very distinguishable in this track. It represents a wind-instrument, probably a trumpet. Immediately after this part, a somewhat longer c01 appears. This part is a very low-pitched string-instrument, it sounds like a string bass. The yellow part of c05 is a very sharp sound, a high-pitched flute which is very unpleasant to the ear. It is clear that both very high- and low pitched sounds are being used at the same time in typical horror music.

Comparison of c03 in both tracks: this sound is in both tracks a very high-pitched instrument, however, in Doll Box this sound is made by a percussion instrument. In Curse Your Name, this sound sounds more like a wind instrument.

Violin plot of timbre coefficients

Analysis performed over the first 60 tracks from each genre

When comparing the twelve Spotify timbre coefficients between the four movie genres, the main difference lays in c03, c06 and c11. C03 is the ‘flatness’ of the sound. A low flatness indicates a “spiky” spectrum (mixture of sine waves) and a high flatness indicates that the spectrum has a similar amount of power in all spectral bands (i.e. similar to white noise). It looks like Action movies have a lower flatness overall, meaning that this genre has a bigger mixture of sounds. Feelgood/Comedy and Horror do have more positive values meaning that these genres have less diversity in spectral bands.

Because timbre features are very hard to interpret it is not possible declare coefficients 6 and 11.

Confusion matrices

Genre classification with all four genres was performed using support vector machines in a ten-fold cross-validation test. Confusion matrices of ‘all features’ (track-level-, timbral- and chroma features), ‘timbral features’ and ‘chroma features’ are being compared. The darker the color grey, the better the prediction. Overall, the chroma features performed the worst among the classification tasks, all features combined predicted movie genre the best.

All features

All features provided the highest genre accuracies for Romance/Drama and Horror. From the confusion matrix, it is clear that there is a clear diagonal dark-grey line, from the upper left to the bottom right. This means, that the model predicted these the best.

Horror and Feelgood/Comedy movies are the most distant when looking at the confusion matrix for all features and timbral features. This makes sense, because Comedies are in general very happy and uptempo, while Horror movie tracks are overall very slow and depressing.

Timbral features

Horror is most often confused with Action movies. This means that the timbral features of Horror movies some what similar are to those of Action movies.

Because the definition of timbral coefficients is somewhat arbitrary, we cannot really conclude what features these exactly are. But when looking at the results of timbre coefficients on the previous chapter (page 7), timbre coefficient c06, c03 and c11 differ across the genres.

Chroma features

Chroma features (pitches) do not seem to predict movie genre very well overall. Music from Action movies is the least distinguishable when looking at chroma features.

From the true Horror tracks, a lot of these tracks are categorized as Action or Feelgood/Comedy. This means that the chroma features from Action look like those of Horror & Feelgood/Comedy movies. It looks like pitches used in Horror and Romantic/Drama tracks are the most distant from each other.

Romantic/Drama tracks differ on chroma features the most. So these songs do have some specific chroma features that most songs contain.

The decision tree is a prediction model that is used in machine learning. The goal is to create a model that predicts the value of a target variable based on several input variables.

From the decision tree classification, the precision and recall for the different genres are much better than the ones for the confusion matrices. The recall for the Horror and Feelgood/Comedy tracks are the best, with a total accuracy of almost 76%. So 76% of the tracks that belong to Horror and Feelgood/Comedy films, are correctly categorized. The prediction is lowest for Action movies.

Forest model

From the forest model, it is very clear that there are some features that really characterize the four movie genres. The top 6 features are:

Track level features

• Valence
• Acousticness
• Danceability
• Instrumentalness

Timbral components

• c06

It seems that timbre c06 is an important component in distinguishing movie genres. However, because timbre components above c04 are very hard to identify, it is not very clear what aspect this exactly is. It looks like valence is the best predictor of all Spotify features.

On the next page, a new analysis on these 5 components is being performed.